Benchmarking AI Token Efficiency Across Programming Languages (Tokenmaxxing)
When prompting AI models, token count directly affects both cost and latency. This benchmark compares how efficiently different languages express similar logic.
Key Findings
Ranking across available tests (Stats Function + Async HTTP).
1. Python : most token-efficient overall. 190 total tokens (Stats: 91, HTTP: 99).
2. JavaScript : 247 total tokens (Stats: 119, HTTP: 128).
3. TypeScript : 274 total tokens (Stats: 137, HTTP: 137).
4. Go : 323 total tokens (Stats: 167, HTTP: 156).
5. Rust : 138 total tokens (HTTP: 138; Stats not measured).
6. Java : 200 total tokens (Stats: 200; HTTP not measured).
Note: Rust and Java were measured in only one of the two tests, so they are not directly comparable to languages with both results.
Python is the most token-efficient language in this benchmark.
Natural Language: English vs Chinese
Pair 1: English 7 tokens vs Chinese 7 tokens (1.0x).
Pair 2: English 9 tokens vs Chinese 12 tokens (1.3x).
Pair 3: English 10 tokens vs Chinese 16 tokens (1.6x).
Pair 4: English 9 tokens vs Chinese 12 tokens (1.3x).
Pair 5: English 12 tokens vs Chinese 19 tokens (1.6x).
Pair 6: English 12 tokens vs Chinese 15 tokens (1.3x).
Pair 7: English 10 tokens vs Chinese 13 tokens (1.3x).
Pair 8: English 9 tokens vs Chinese 13 tokens (1.4x).
Across these samples, Chinese generally uses more tokens than English (about 1.3x to 1.6x).
Why this happens
Many LLM tokenizers are more optimized for English subwords and common word fragments. Chinese text is often split into smaller units, which can increase token count for equivalent meaning.
Methodology
Tested real implementations (stats functions, async HTTP). Measured token count with OpenAI tokenizer. Languages tested: Python, JavaScript, TypeScript, Go, Rust, Java, C, C++. Also tested natural language samples in English and Chinese.
What's Next?
Looking at scripting languages (Ruby, Perl, Lua), markup (HTML, YAML, XML), styling (CSS, SCSS), and queries (SQL, GraphQL).